One-line summary
Neon scores 61 at $3/$15. Sol scores 57 at $2/$10. Astra scores 62 at $8/$40. Three price tiers for three capability tiers — the dividing line is not the score, it is how much error the task tolerates.
Why these three belong together
Between November 17 and 19, 2026, two vendors shipped three models into the same band. OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17; Google released Gemini 4 Neon on November 19. None is a flagship; all three occupy the transition zone between daily workhorse and complex-task specialist.
They are comparable for three reasons: released in the same week, measured by the same third-party index, and billed per token.
Core comparison
| Dimension | Gemini 4 Neon | GPT-6.2 Sol | GPT-6.2 Astra |
|---|---|---|---|
| Vendor | OpenAI | OpenAI | |
| Release | 2026-11-19 | 2026-11-17 | 2026-11-17 |
| Intelligence index | 61 (third globally) | 57 | 62 (second globally) |
| Input per 1M tokens | $3 | $2 | $8 |
| Cached input per 1M | $0.15 | $0.10 | $0.40 |
| Output per 1M tokens | $15 | $10 | $40 |
| Context window | 1M tokens | 1.05M tokens | 1.05M tokens |
| Max output | not separately listed | 130K tokens | 130K tokens |
| Adjustable reasoning depth | four levels | not separately listed | not separately listed |
| Positioning | complex coding, analysis, enterprise workflows | daily work, coding, multi-step agents | complex professional work, flagship reasoning |
| Availability | AI Studio, Gemini app, enterprise API, Antigravity | ChatGPT, API, Codex | ChatGPT, API, Codex |
Prices are standard processing rates; long-context and batch tiers may be priced separately. Index scores are third-party measurements and may use a different index version than this site's October Carbon dataset, so cross-version comparisons are not valid.
The arithmetic behind three price tiers
Put price next to score and three distinct curves appear.
Per million input tokens, Neon costs 50% more than Sol ($3 vs $2) for four points more (61 vs 57). Astra costs 300% more than Sol ($8 vs $2) and 167% more than Neon ($8 vs $3) for one point over Neon (62 vs 61) and five over Sol.
On the output side the gaps widen: Neon $15, Sol $10, Astra $40. Astra's output price is 2.67x Neon and 4x Sol.
The implication: Astra's premium buys reliability on complex professional tasks, not across-the-board intelligence. Between 61 and 62 the difference is a single point at 2.67x the price. Whether that point is worth it depends on whether the task tolerates failure — in long-horizon autonomous work, one failed run can cost far more than the tokens.
Three workload profiles
High-volume, short tasks, human review in the loop. Batch classification, summarization, structured extraction. Single-run accuracy matters less than unit cost; Sol at $2/$10 is the lowest price, and its 1.05M context can absorb long documents in one pass.
Complex coding and long-document analysis. Cross-file refactors, long-report reconciliation, heterogeneous-material consolidation. Neon's 1M context, four reasoning levels and $3/$15 price make it the only mid-priced high-context option; Sol is cheaper but four points lower, which raises rework odds on complex refactors.
Multi-step autonomous execution where errors are expensive. Long-horizon agent work and professional workflows. Astra's $8/$40 buys 62 and greater autonomy, but note that OpenAI's own early testers complained Astra asked for permission too often and struggled with autonomous workflows. That should be validated independently rather than inferred from the score.
Two variables that often get missed
The first is cached input pricing. The spread is wider than headline rates: Sol $0.10, Neon $0.15, Astra $0.40. For workflows that repeatedly query the same long document, cache hit rates determine whether the actual bill is a fraction of the sticker price — a 500K-token document queried repeatedly makes cache economics the dominant cost driver.
The second is speed. Google states Neon runs about 3x the speed of Carbon at 60% of its cost; OpenAI states Astra is the first public model assessed as reaching the "High" threshold for autonomous improvement. These measure different things — inference throughput versus self-improvement — and cannot be converted into "who is faster."
Verdict
If you rank purely by score divided by price, Sol wins, Neon is second and Astra is clearly expensive.
Real selection does not work that way. Fix the error tolerance the task allows, then pick the cheapest model that clears that bar:
- Human review, repairable errors → Sol;
- Must be right the first time, long context, mid-priced → Neon;
- Multi-step autonomy, expensive failure → Astra, validated on a real task first.
Under that logic, "61 or 57" is not the decision. "How many failures does this task allow" is.
Boundary
- Release details, pricing and context windows come from vendor pages and reporting by Cailianshe and 36Kr;
- Artificial Analysis scores are third-party and not comparable across index versions;
- Speed and autonomy claims are vendor statements without disclosed methodology;
- This article does not evaluate real-world output quality; selection guidance is cost-structure inference, not procurement advice;
- Long-context pricing and batch discounts follow each vendor's official price page.
